Concurrency

OrientDB uses an optimistic approach to concurrency. Optimistic Concurrency Control, or OCC assumes that multiple transactions can compete frequently without interfering with each other. It's very important that you don't share instances of databases, graphs, records, documents, vertices and edges between threads because they are non thread-safe. For more information look at Multi-Threading.

How does it work?

Consider the following scenario, where 2 clients, A and B, want to update the amount of a bank account:

       Client A                               Client B

           |                                     |
          (t1)                                   |
   Read record #13:22                            |
     amount is 100                              (t2)
           |                             Read record #13:22                   
          (t3)                             amount is 100
  Update record #13:22                           |
set amount = amount + 10                        (t4)
           |                            Update record #13:22
           |                           set amount = amount + 10
           |                                     |

Client A (t1) and B (t2) read the record #13:22 and both receive the last amount as USD 100. Client A updates the amount by adding USD 10 (t3), then the Client B is trying to do the same thing: updates the amount by adding USD 10. Here is the problem: Client B is doing an operation based on current information: the amount was USD 100. But at the moment of update, such information is changed (by Client A on t3), so the amount is USD 110 in the database. Should the update succeed by setting the new amount to USD 120?

In some cases this could be totally fine, in others not. It depends by the use case. For example, in your application there could be a logic where you are donating USD 10 to all the accounts where the amount is <=100. The owner of the account behind the record #13:22 is more lucky than the others, because it receives the donation even if it has USD 110 at that moment.

For this reason in OrientDB when this situation happens a OConcurrentModificationException exception is thrown, so the application can manage it properly. Usually the 3 most common strategies to handle this exceptions are:

1. Retry doing the same operation by reloading the record #13:22 first with the updated amount
2. Ignore the change, because the basic condition is changed
3. Propagate the exception to the user, so he can decide what to do in this case

Optimistic Concurrency in OrientDB

Optimistic concurrency control is used in environments with low data contention. That is, where conflicts are rare and transactions can complete without the expense of managing locks and without having transactions wait for locks to clear. This means a reduced throughput over other concurrency control methods.

OrientDB uses OCC for both Atomic Operations and Transactions.

Atomic Operations

OrientDB supports Multi-Version Concurrency Control, or MVCC, with atomic operations. This allows it to avoid locking server side resources. At the same time, it checks the version in the database. If the version is equal to the record version contained in the operation, the operation is successful. If the version found is higher than the record version contained in the operation, then another thread or user has already updated the same record. In this case, OrientDB generates an OConcurrentModificationException exception.

Given that behavior of this kind is normal on systems that use optimistic concurrency control, developers need to write concurrency-proof code. Under this design, the application retries transactions x times before reporting the error. It does this by catching the exception, reloading the affected records and attempting to update them again. For example, consider the code for saving a document,

int maxRetries = 10;
List<ODocument> result = db.query("SELECT FROM Client WHERE id = '39w39D32d2d'");
ODocument address = result.get(0);

for (int retry = 0; retry < maxRetries; ++retry) {
     try {
          // LOOKUP FOR THE INVOICE VERTEX
          address.field( "street", street );
          address.field( "zip", zip );
          address.field( "city", cityName );
          address.field( "country", countryName );

          address.save();

          // EXIT FROM RETRY LOOP
          break;
     }
     catch( ONeedRetryException e ) {
          // IF SOMEONE UPDATES THE ADDRESS DOCUMENT
          // AT THE SAME TIME, RETRY IT.
     }
}

Transactions

OrientDB supports optimistic transactions. The database does not use locks when transactions are running, but when the transaction commits, each record (document or graph element) version is checked to see if there have been updates from another client. For this reason, you need to code your applications to be concurrency-proof.

Optimistic concurrency requires that you retire the transaction in the event of conflicts. For example, consider a case where you want to connect a new vertex to an existing vertex:

int maxRetries = 10;
for (int retry = 0; retry < maxRetries; ++retry) {
     try {
          // LOOKUP FOR THE INVOICE VERTEX
          Vertex invoice = graph.getVertices("invoiceId", 2323);

          // CREATE A NEW ITEM
          Vertex invoiceItem = graph.addVertex("class:InvoiceItem");
          invoiceItem.field("price", 1000);

          // ADD IT TO THE INVOICE
          invoice.addEdge(invoiceItem);

          graph.commit();

          // EXIT FROM RETRY LOOP
          break;
     }
     catch( OConcurrentModificationException e ) {
          // SOMEONE HAS UPDATED THE INVOICE VERTEX
          // AT THE SAME TIME, RETRY IT
     }
}

Concurrency Level

Before v2.2.4, transactions acquire an exclusive lock on the storage, so no matter if you have 1 or 100 cores, the execution was always serialized. With 2.2.4 and further, transactions are executed in parallel only if they involve different clusters and indexes.

In order to use the transaction parallelism, the domain has to be slightly changed by using the OrientDB inheritance by creating a base class and multiple sub-classes, one per core. Example of creating the class Log with 4 sub-classes (4 cores) and the indexed property 'id':

CREATE CLASS Log ABSTRACT

CREATE CLASS Log_1 EXTENDS Log
CREATE PROPERTY Log_1.id STRING
CREATE INDEX Log1Id ON Log_1(id) UNIQUE_HASHINDEX

CREATE CLASS Log_2 EXTENDS Log
CREATE PROPERTY Log_2.id STRING
CREATE INDEX Log1Id ON Log_2(id) UNIQUE_HASHINDEX

CREATE CLASS Log_3 EXTENDS Log
CREATE PROPERTY Log_3.id STRING
CREATE INDEX Log1Id ON Log_3(id) UNIQUE_HASHINDEX

CREATE CLASS Log_4 EXTENDS Log
CREATE PROPERTY Log_4.id STRING
CREATE INDEX Log1Id ON Log_4(id) UNIQUE_HASHINDEX

After creating multiple sub-classes, you should bind your threads/client (it depends, respectively, if you are working in embedded mode or client/server) to a different sub-class. For example with 4 cores, you have 4 sub-classes (like above) and this could be the binding for the class "Log":

• Thread/Client 1 -> Class Log_1
• Thread/Client 2 -> Class Log_2
• Thread/Client 3 -> Class Log_3
• Thread/Client 4 -> Class Log_4

If you are working with graphs, it's a good practice to apply the same rule to both vertex and edge classes. In this example we have 4 cores, so 4 clusters per vertex class and 4 clusters per edge class:

• Thread/Client 1 -> Classes User_1 and City_1 for vertices and Class Born_1 for edges
• Thread/Client 2 -> Classes User_2 and City_2 for vertices and Class Born_2 for edges
• Thread/Client 3 -> Classes User_3 and City_3 for vertices and Class Born_3 for edges
• Thread/Client 4 -> Classes User_4 and City_4 for vertices and Class Born_4 for edges

Now look at these 2 SQL scripts:

Client 1:

BEGIN
LET v1 = CREATE VERTEX User_1 SET name = 'Luca'
LET v2 = CREATE VERTEX City_1 SET name = 'Rome'
CREATE EDGE Born_1 FROM $v1 TO $v2
COMMIT RETRY 10

Client 2:

BEGIN
LET v1 = CREATE VERTEX User_2 SET name = 'Luca'
LET v2 = CREATE VERTEX City_2 SET name = 'Rome'
CREATE EDGE Born_2 FROM $v1 TO $v2
COMMIT RETRY 10

In this case the two transactions go in parallel with no conflict, because they work on different classes and indexes.

Thanks to the OrientDB polymorphism, sub-classes are instance of the abstract class, so you can still execute queries by using the base class as target and OrientDB will consider all the sub-classes, so your model remains clean at application level. Example:

SELECT * FROM User WHERE name = 'Luca'

But if you already know that Luca if exists is in the 2nd partition of the User class (User_2 sub class), you can also execute:

SELECT * FROM User_2 WHERE name = 'Luca'

When it's possible to pre-determine there the record is saved, using the sub-class as target has better performance.

Concurrency when Adding Edges

Consider the case where multiple clients attempt to add edges on the same vertex. OrientDB could throw the OConcurrentModificationException exception. This occurs because collections of edges are kept on vertices, meaning that, every time OrientDB adds or removes an edge, both vertices update and their versions increment. You can avoid this issue by using RIDBAG Bonsai structure, which are never embedded, so the edge never updates the vertices.

To use this configuration at run-time, before launching OrientDB, use this code:

OGlobalConfiguration.RID_BAG_EMBEDDED_TO_SBTREEBONSAI_THRESHOLD.setValue(-1);

Alternatively, you can set a parameter for the Java virtual-machine on startup, or even at run-time, before OrientDB is used:

$ java -DridBag.embeddedToSbtreeBonsaiThreshold=-1

	While running in distributed mode SBTrees are not supported. If using a distributed database then you must set ridBag.embeddedToSbtreeBonsaiThreshold = Integer.MAX\_VALUE to avoid replication errors.

Troubleshooting

Reduce Transaction Size

On occasion, OrientDB throws the OConcurrentModificationException exception even when you concurrently update the first element. In particularly large transactions, where you have thousands of records involved in a transaction, one changed record is enough to roll the entire process back with an OConcurrentModificationException exception.

To avoid issues of this kind, if you plan to update many elements in the same transaction with high-concurrency on the same vertices, a best practice is to reduce the transaction size.